search

bruinxiong / Modified-CRUNet-and-Residual-Attention-Network.mxnet

:fire::fire:A MXNet implementation of Modified CRUNet & Residual Attention Network:fire::fire:
Apache License 2.0
67 stars 30 forks source link

results on Residual-Attention-Network #4

Open rexnxiaobai opened 6 years ago

rexnxiaobai commented 6 years ago

how about the experiment details and results on Residual-Attention-Network?

bruinxiong commented 6 years ago

@rexnxiaobai Thank you for your attention. Initially, we can not obtain the performance of verification (98.48%) mentioned in arcface paper on r100 model for megaface challenge 1, we only get 98.087% under our trained model on MS1M dataset and data preprocessing. After we deployed Residual attention network trained on emore dataset, we can obtain the 98.355% performance of verification. As a comparison, we also evaluate r100 trained on emore dataset, the performance of verification is 98.111%.

billtiger commented 4 years ago

@bruinxiong
the 98.355% performance is acheived by Residual attention network 56 or 92?